IBM Releases Two Experimental AI Models That Separate Instructions From Source Text to Resist Hidden Attacks

At a glance

IBM has released two experimental AI models that read instructions and reference material separately. The goal is to keep hidden malicious commands in documents or tool outputs from changing what the AI is supposed to do. The models were published on the Hugging Face technical blog.

Key points

• The two models are 3B and 8B versions based on Granite 4.1.
• They were trained to read and generate LLMON, a format that splits input into instruction, reference, and output-condition blocks.
• IBM says they resist malicious instructions better in several tests, but defenses against attacks that change over multiple conversations are still being improved.
• The 3B model supports 12 languages, including Korean, and is released under the Apache 2.0 license.

Why it matters

Keeping AI from being steered by hidden commands in documents and external tools can make business AI assistants and agents safer to use.

More from this field

• GitHub Next Releases Ace, a Mac Coding Space Where Teams and AI Agents Work Together

  GitHub Next has released Ace, an experimental coding environment on Mac where team members and AI agents code together. Teams discuss in a channel, then call an agent to take on the work.

  GeekNews · Nov 14, 09:30 KST
• Claude Whiteboards, Sheets and Data: What Works Now

  This video tries out Claude's new whiteboard, sheet (spreadsheet), and data features to see which ones are actually useful. It also asks whether there is a reason to move from Google Sheets.

  The AI Advantage · Nov 14, 08:57 KST
• Amazon Plans Index of Online Product Data Across the Web

  Amazon reportedly plans to gather data on products sold anywhere online, including items it does not sell, and show it on its own site, according to Reuters. The project would cover products beyond Amazon's own catalog.

  The Verge · Nov 14, 08:49 KST
• Theo Retracts Praise for Gemini 4 Carbon Over Severe Response Delays

  This video is from the Theo (t3.gg) channel. Theo corrects his earlier mild review of Gemini 4 Carbon and covers the response delays he found in real-world use. Gemini 4 Carbon is an AI model from Google.

  Theo - t3.gg · Nov 14, 08:45 KST